Simulating clinical interventions with a generative multimodal model of human physiology
Guy Lutsker, Gal Sapir, Jordi Merino, Smadar Shilo, Anastasia Godneva, Eli Meirom, Shie Mannor, Hagai Rossman, Gal Chechik, Eran Segal
cs.AI
2026-04-30
A 139M decoder-only model trained on 15k phenotyped people matches 41 RCT directions zero-shot and predicts individual 6-month diastolic BP change at r=0.78.
Clinical decisions depend on how physiology changes over time, and why the same intervention lands differently across people. Most clinical AI still fits the medical record: diagnosis codes, orders, sparse labs. That record is a log of encounters. Diet, sleep, adiposity, continuous glucose, and the gut microbiome, the quantities that move health between visits, barely appear. Care-derived cohorts also over-represent symptomatic people, so pre-clinical trajectories are hard to learn.
Single-modality foundation models are strong inside one signal, such as retinal images, CGM, or clinical text. Task-specific effect models answer one endpoint at a time, statins on LDL or exercise on blood pressure. Static multimodal fusion can stack features. It does not give multi-year dynamics.
The missing object is a time-ordered physiological trajectory that can be queried for a future value or a hypothetical intervention. HealthFormer is trained for that object.
Training data come from the Human Phenotype Project, led at the Weizmann Institute: 15,319 deeply phenotyped people, split 12,102 / 3,217. Among test participants, 1,364 have a second visit about two years later. The corpus holds about 64.2 million tokens. Each person becomes a chronological sequence of up to 25,000 tokens covering 667 measurements in 19 groups (blood, DEXA, sleep, CGM, microbiome, wearables, medications, diet, ECG, ultrasound, and others). Continuous variables are quantile-binned on the training set (roughly 15-20 bins for blood tests, 129 for CGM); categoricals are enumerated. The vocabulary has 13,056 tokens. Each position sums six embeddings: discrete token, raw continuous value, modality, a 7-dimensional clock (weekday, hour, month, year, and related fields), sinusoidal position, and age/sex.
The backbone is a 139-million-parameter decoder-only transformer: 14 layers, width 768, 12 heads, trained on 8 H100 GPUs with an effective batch of 16. A query block injects the target modality and timestamp into the last hidden state, so one forward pass can request any measurement at any time. Interventions are not a separate head. Drugs and exercise are extra categorical tokens appended to the sequence; token count and spacing encode frequency and duration. Diet and CPAP change existing continuous values and re-bin them. The loss adds binned cross-entropy, z-scored MAE, and a split-context term that is bidirectional inside visit 1 and causal from visit 1 into later visits. Missing assays are simply absent from the sequence.
Within a visit, category-level Pearson r is 0.82 for ankle-brachial index, 0.81 for DEXA, and 0.79 for sleep; some single biomarkers exceed 0.9. Forecasting visit 2 from visit 1, 330 of 338 modalities are significant. Anthropometry averages r = 0.65, the microbiome 0.64, ultrasound 0.25. Against per-modality linear regression, HealthFormer wins 51 of 57 modalities with a significant difference. The mean lift across all modalities is only Δr = +0.001, a thin gain over last-value plus age, sex, and BMI. Scaling from 14M to 139M parameters raises longitudinal r monotonically.
With frozen weights and tokenizer, zero-shot transfer reaches a median r of 0.52 on UK Biobank (71 of 74 shared modalities significant), 0.39 on NHANES, 0.45 on Framingham, and 0.70 on PNP3. Mean-pooled UK Biobank embeddings plus a penalized Cox model beat age + sex + BMI on 27 of 30 incident-disease and mortality endpoints. The largest lifts are COPD (+0.13), chronic kidney disease (+0.09), and hypertension (+0.07). Absolute C-index is 0.826 for cardiovascular death, 0.808 for heart failure, 0.834 for CKD. Where clinical scores exist, HealthFormer is higher in every comparison: +0.06 over Framingham CVD for hypertension, +0.17 over PREVENT-ASCVD for ischaemic stroke.
The distinctive claim is intervention conditioning. PNP3's 200 participants were held out. Given baseline labs, six months of CGM, and diet logs, predicted six-month change correlates with observed change at r = 0.78 for diastolic blood pressure (n = 194), 0.63 for BMI, 0.61 for fasting glucose, 0.89 for triglycerides, 0.49 for HbA1c. HDL is not significant (n = 15). Mean r on six well-powered biomarkers rises from 0.39 with baseline only to 0.64 with full context. Mean absolute error beats a no-change baseline on 9 of 10 biomarkers; HDL is the miss.
Against 41 published RCT comparisons, using synthetic n = 200 cohorts matched to each trial's Table 1, predicted direction agrees in 41 of 41 and the predicted mean falls inside the published 95% CI in 30 of 41. Recovered endpoints include ALLHAT amlodipine and chlorthalidone on systolic pressure, ezetimibe on LDL, and most SGLT2, metformin, and exercise glycaemic endpoints. Misses cluster. High-intensity statins and high-dose semaglutide in STEP-1 are under-predicted (rosuvastatin on LDL 24.7% vs 46% published; weight 8.3% vs 14.9%). Ramipril systolic change in HOPE is over-predicted (10.1% vs 2.4%). Off-target blood-pressure and weight effects of SGLT2 and GLP-1 agents run 2-9 points high.
| Comparison | Metric | Result |
| Within-visit ABI / DEXA / sleep | category-mean Pearson r | 0.82 / 0.81 / 0.79 |
| Two-year forecast vs per-modality linear | significant wins; overall Δr | 51/57; +0.001 |
| UKB 30 endpoints vs age+sex+BMI | improved endpoints | 27/30 |
| PNP3 six-month diastolic BP change | individual Pearson r | 0.78 (n=194) |
| 41 RCTs | direction; mean inside published 95% CI | 41/41; 30/41 |
For anyone building clinical models or digital twins, this is a full swap of the modelling object, from the medical record to a physiological trajectory. Forecasting, risk, and intervention queries share one generative setup. At 139M parameters on 8 H100s, the recipe is in reach.
The usable surface is still narrow. Cross-cohort reconstruction and risk embeddings can be tried as features. Drug choice cannot be treated as a counterfactual engine. The paper's own label is an initial health world model, not a decision-grade clinical digital twin.
The average two-year gain over linear regression is near zero. What the unified weights add is intervention conditioning, and transfer onto UK Biobank, where only 74 measurements overlap.
Training is observational. The model learns co-occurrence, not causation. Intervention queries compare exposed versus unexposed trajectories. They do not estimate what would happen if the same person switched drugs, which would need multiple comparable interventions per person; HPP does not provide that. The RCT panel uses synthetic cohorts, not original trial participants, and only continuous biomarkers, not hard events.
Dose, adherence, titration, and discontinuation remain coarse. Statin tokens do not separate moderate from high intensity. Recent high-dose GLP-1 exposure is sparse in HPP, so large-effect tails regress toward the smaller changes that dominate a general-population cohort. PNP3 was held out at the person level, but it shares the same group, ascertainment, CGM device, and diet-logging protocol as HPP, which inflates the individual correlations.
On the longitudinal task, linear regression trails by 0.001 in mean r. Cross-system gain shows up mainly in the subset with significant differences. Disease prediction uses pooled embeddings plus Cox, not the generative query itself. RCT inclusion also required the intervention to appear at least 30 times in training and the local biomarker r to be at least 0.30, which drops endpoints the model already cannot predict.